ODOR-SIGHT
Predict odorant molecules
with explainable AI
Deep learning web server using Graph Attention Networks to classify molecules as odorant or odorless with 94%+ accuracy and bond‑level explanations.
ODOR-SIGHT
Predict odorant molecules with explainable AI
Deep learning web server using Graph Attention Networks to classify molecules as odorant or odorless with 94%+ accuracy and bond‑level explanations.
The architecture
ODOR-SIGHT represents molecules as graphs — atoms as nodes, bonds as edges — and applies multi-head attention to capture complex structural patterns relevant to olfaction. Each atom encodes 45 features including atomic properties and physicochemical descriptors; each bond encodes 11 features capturing bond type, stereochemistry, and conjugation.
Interpretability
EdgeSHAPer provides bond-level Shapley value explanations, revealing which chemical bonds drive each prediction. The model-coupled applicability domain measures maximum layer-3 atom novelty against a leave-one-molecule-out p95 boundary, flagging unfamiliar atom environments for additional scrutiny.
GRAPH ATTENTION NETWORK
ODOR-SIGHT uses a Graph Attention Network (GAT) where molecules are graphs of atoms and bonds. Nodes encode atomic properties (type, hybridization, aromaticity) and descriptors like LogP and Gasteiger charges. Multi-head attention dynamically weights atom neighborhoods for rich structural learning.
HIGH ACCURACY
Powered by an optimized GATv2 with 4 layers, ODOR-SIGHT exceeds 0.94 Balanced Accuracy (BACC) on external validation. The model learns from 45 node and 11 edge features, rigorously validated on independent test sets to ensure reliable generalization.
CURATED OLFACTION DATA
Our training dataset comprises carefully curated olfaction data from established databases. Each compound is validated for odorant or odorless classification, providing a robust foundation for model training and reliable predictions.
EDGESHAPER EXPLANATIONS
EdgeSHAPer explains predictions by quantifying each bond's contribution via Shapley values. View bond importance heatmaps, identify key influential bonds, and explore pertinent positive sets—the smallest subsets sufficient to maintain the prediction.
APPLICABILITY DOMAIN
ODOR-SIGHT assesses reliability with model-coupled atom-level novelty from GNN layer 3. Each atom is compared with training atom environments, and the most novel atom defines the molecular score. A molecule-level leave-one-out p95 boundary determines the official In/Out status; the PCA map is retained only as visual context.
OECD COMPLIANT
ODOR-SIGHT follows OECD QSAR guidelines, including defined endpoints, unambiguous algorithms, applicability domain characterization, and mechanistic interpretation, ensuring robust and regulatory-ready predictions.
Odor-Sight · Predictor
Instructions
Draw a substance or paste a smiles string
Draw the chemical structure of the compound you want to evaluate using the drawing tool provided in the app. Alternatively, you can paste a SMILES string of the compound into the drawer. You can review and edit it if necessary.
Or upload a single SDF / MOL file
Use the editor's load button to import one structure from an SDF or MOL file (max 2 MB). The editor reads the file in your browser and loads a single molecule onto the canvas. ODOR-SIGHT predicts one molecule per submission and shows the server-side curation for confirmation before prediction.
Review the curated input and confirm
Click 'REVIEW MOLECULE' to see the original structure, the curated structure and every transformation. Confirm the exact form that should enter the prediction queue.
Review the results and download the report
After the prediction is complete, you will see the results including the classification (odorant or odorless), confidence score, applicability domain analysis, and EdgeSHAPer explanations. You can download a detailed PDF report.
Before prediction, the server checks the structure, removes identifiable salts and counter-ions, standardizes the representation, neutralizes charges where appropriate and selects a canonical tautomer. You will see and confirm every change first. Submitted structures are not stored as readable application logs.